This article is more than 5 years old. Information may no longer be current.
Smartphone heart rate tech, deep learning may aid in diabetes detection
Click Here to Manage Email Alerts
NEW ORLEANS — A smartphone application that tracks a person’s heart rate may be able to detect diabetes using a photoplethysmography signal, which is easily measured using a smartphone’s light and camera, according to a study presented at the American College of Cardiology Scientific Session.
“We’ve demonstrated that by using deep learning and a smartphone camera alone, we can also detect vascular changes associated with diabetes and with reasonable discrimination,” Robert Avram, MD, postdoctoral fellow at UCSF Medical Center, said in a press release.
Avram and colleagues analyzed data from nearly 55,000 participants enrolled in the Health eHeart study who used a smartphone app to track their heart rate (Azumio Instant Heart Rate).
“While analyzing the heart rate data as collected using smartphone apps in the Health eHeart study, we noticed that patients with diabetes had, on average, a higher ‘free-living’ heart rate than patients without diabetes when adjusted for multiple factors,” Avram told Cardiology Today.
“This pushed us to analyze the photoplethysmography signal to see if there were other features that would help differentiate patients with diabetes from patients without diabetes. By identifying these features, we saw a huge opportunity to develop a screening tool for diabetes using deep learning and a smartphone camera and flash to classify patients as having prevalent diabetes or no diabetes.”
Variations in blood volume that occur with every heart beat can be captured by shining a smartphone flashlight on a fingertip, Avram said in a press release. With every heart contraction, BP increases in the vessels, causing them to expand, which then increases the amount of light reflected by the skin to the optical sensor of the phone’s camera. This input can then be converted to a waveform representing the volumetric change of the blood volume in a vessel.
High detection rate
The mean age of the participants in the study was 45 years, 53% were men and 7% had self-reported diabetes.
The researchers randomly divided participants into separate training, development and test data sets. The development data set was used for model tuning, and model discrimination was measured using area under the receiver-operating characteristic curves in the test data set. Avram and colleagues developed and applied a deep learning algorithm that used the smartphone-based photoplethysmography signal recordings to identify which patients had diabetes based on this signal alone.
The model correctly identified diabetes in 72% of cases using photoplethysmography only, with a negative predictive value of 97%, according to results presented here.
“Using smartphone-based photoplethysmography, we can detect prevalent diabetes in a large ambulatory sample of nearly 3 million recordings with reasonable discrimination. Once further validation is conducted in minorities and in a prospective, in-clinic cohort, we will implement our deep learning model in the app. Users of the [Azumio] Instant Heart Rate app could use it to screen for prevalent diabetes and then follow up with a physician for a confirmation of the diagnosis. The development of screening tools that can be easily deployed in free-living settings could increase access to diabetes screening beyond the reach of the traditional clinic’s ‘sphere of influence,’ to identify more at-risk individuals and, ultimately, decrease the prevalence of undiagnosed diabetes,” Avram told Cardiology Today.
Notably, after combining the model with other common risk factors for diabetes, the model’s ability to detect diabetes increased to 81%.
“I was surprised to find out that after coupling the app-based screening with common risk factors — age, sex, ethnicity and BMI — our tool was comparable to many traditional diabetes risk scores that are used in clinics to predict diabetes,” Avram said, noting that this creates new avenues for improving care. “Our tool can be self-administrated by the user, without requiring a physician for assistance.”
Expanding research
Despite these promising findings observed in this study, gaps in knowledge remain and some areas require further study, according to Avram.
“Currently, we validated our algorithm in nearly 60,000 users where the diagnosis of diabetes was self-reported. While self-reported diagnoses in the Health eHeart study have previously been found to be valid when compared to electronic medical record data, we would like to validate our algorithm prospectively, using the electronic medical record diabetes diagnosis, rather than the self-reported diabetes diagnosis,” he said.
Avram and colleagues are currently conducting such validation in two CV prevention clinics.
“Moreover, we need to explore the performance of our algorithm in other minorities who are at higher risk from suffering from undiagnosed diabetes, such as African-Americans and Asians,” he told Cardiology Today. – by Melissa Foster
Reference:
Avram RA. Abstract 1276-422. Presented at: American College of Cardiology Scientific Session; March 16-18, 2019; New Orleans.
Disclosure: Avram reports no relevant financial disclosures.